Outliers can be caused by various factors like data entry errors, measurement errors, and data processing errors. For example, entering an extra digit while recording data or measurement device calibration errors.

CDA typically comes after the EDA stage in the data analysis process. EDA allows analysts to explore the data and generate hypotheses while CDA involves statistical tests to confirm or refute these hypotheses.

In a uniform distribution, all values occur with the same frequency, so the "Mean", "Median", and "Mode" would all be the same, falling in the center of the distribution.

In the context of your analysis, a feature with negligible variance might have little influence on the outcome variable. This is because, with very little variance, the feature is nearly constant and hence, provides no new information for the model. Depending on the context and the objectives of your analysis, you might consider dropping this feature.

A Histogram is often used to understand the underlying frequency distribution of data. It groups numbers into ranges (bins) and the height of each bar depicts the number of values that fall into each range.

An outlier might not be visible in a box plot if the outlier is close to the whisker because it might still fall within the range of the whisker and thus be considered part of the normal distribution.

Understanding the audience's knowledge and familiarity with different types of graphs can significantly influence the choice of graph. For a general audience, simpler graphs like bar charts or line graphs may be more suitable, whereas a more technical audience might be comfortable with more complex visualizations like heatmaps or network diagrams.

Multicollinearity affects the coefficients of a linear regression model by making them unstable (small changes in the data cause large swings in the coefficients), inflating the standard errors of the coefficients (making them less statistically significant), and weakening the statistical power of the model (decreasing the chances of finding valid effects).

Wrapper methods for feature selection assess the relevance of a subset of features by considering their ability to predict the outcome with a particular learning algorithm. These methods involve training a model multiple times and selecting the subset of features that maximizes the model's performance.

Bias might be introduced in model-based imputation if the chosen model fits poorly to the data. If the model used does not reflect the true data generation process, the imputed values might be systematically biased, leading to incorrect conclusions.